A GitHub repository called T3 Code added roughly 1,150 stars today, and its README opens by answering the question its own kind usually dodges: "Wait, what are you selling me?" The answer is nothing.
T3 Code, from the streamer and tooling developer Theo Browne's ping.gg, calls itself an "agent harness control surface." That phrase is doing real work. It is not a coding agent and it ships neither a model nor a subscription. It is a remote control for the agents already installed on your machine: Claude Code, Codex, Cursor's agent CLI, Grok Build, OpenCode and Google Antigravity. If a provider is set up and authenticated locally, T3 can drive it.
The practical difference is where the interface lives. There is an iOS app, an Android app, a local web app served by a `t3` command, and an Electron desktop build. The server runs on your machine, so a phone on the same network — or reached over a tunnel — can watch and steer a session that is editing files on your desk. For anyone who has started a long agent run and then gone to make coffee, that is the feature.
The install path is deliberately low-friction: `curl -fsSL https://t3.codes/install.sh | sh`, or on Windows `irm https://t3.codes/install.ps1 | iex`; plus `npx t3@latest`, winget, scoop, a Homebrew cask, a .deb package and an AUR package with a nightly channel. `t3 service install` turns the server into a background service and `t3 update` pulls newer releases.
What is unusual is how candid the project is about its own state. "We are very very early in this project. Expect bugs." It is mostly not accepting contributions — small fixes maybe, big features no — and there is no documentation site, only a `docs/` directory. Building from source needs Vite's `vp` tool, which tells you something about the age of the stack.
What it does have is the plumbing a control surface needs: permission modes, keyboard shortcuts, per-project settings, appearance preferences, multi-account support for Codex and Claude, source-control integrations, MCP connections so app and server stay in sync, and remote access documented as a first-class feature rather than a hack.
The strategic bet is worth naming. For two years the assumption has been that developers standardise on one harness and one vendor. T3 Code assumes the opposite: that a working developer in 2026 may hold a Claude subscription, a Codex subscription and something else besides, and that the layer which remembers all of them is where the daily habit forms. Whoever owns the control surface does not have to own the model.
That is also why the "nothing" answer matters. Codex has a desktop app, Conductor and Claude Desktop exist, Cursor shipped Glass — and the README says none of them met the bar on performance, remote-readiness and openness. Publishing the code under MIT with an explicit invitation to fork it is an argument that this layer should not be a moat.
Two cautions. First, a trending repository measures attention, not adoption; 1,148 stars in a day say many developers recognised the problem, not that any of them installed the thing. Second, a control surface adds a failure mode: when the GUI is wrong, you are debugging the agent and the harness at once. The most useful line in the README remains that T3 does not run its own AI. If the agent misbehaves, the agent is still the suspect.
